Treat retrieved instructions as untrusted content

Illustrative editorial sample. This draft describes a general workflow, not a current announcement, quotation or measured benchmark. It requires editorial review before publication.

A practical starting point

An illustrative security review for systems that read documents and use tools.

1. Define the boundary

Identify the trust boundary between application instructions and retrieved material. A document should not grant new permissions.

2. Inspect the evidence

Authorize each tool call using application identity and policy. Do not rely on the model to enforce access control.

3. Make the decision explicit

Redact secrets before context construction. Include only resources the current user is permitted to inspect.

4. Test the failure path

Test hostile text, misleading links, malformed outputs and attempts to access another account. Verify rejection without leaking private context.

Review before use

Adapt this checklist to your application, record the assumptions and verify any product-specific facts against primary sources. This sample does not provide a news report, legal opinion or performance guarantee.